[Relay] Convert negative axes to positive when importing ONNX Unsqueeze

python/tvm/relay/frontend/onnx.py

@@ -2540,6 +2540,8 @@ class Unsqueeze(OnnxOpConverter):
     def run_calculation(cls, tensor, axes):
         axes = sorted(axes)
         for axis in axes:
+            if axis < 0 and isinstance(tensor, _expr.Var):
+                axis = len(tensor.type_annotation.concrete_shape) + len(axes) + axis
             tensor = _op.expand_dims(tensor, axis=axis, num_newaxis=1)
         return tensor
 
@@ -2558,6 +2560,7 @@ def _impl_v13(cls, inputs, attr, params):
         num_new_axis = int(infer_type(inputs[1]).checked_type.shape[0])
         axes = relay.sort(inputs[1])
         axes = relay.split(axes, num_new_axis).astuple()
+        rank_output = rank_input + num_new_axis
         result = inputs[0]
 
         # TODO (AndrewZhaoLuo): investigate performance issues with consecutive
@@ -2567,7 +2570,7 @@ def _impl_v13(cls, inputs, attr, params):
             # Unpack scalar
             axis = relay.reshape(axis, [])
             axis = relay.where(
-                axis >= relay.const(0, "int64"), axis, axis + relay.const(rank_input, "int64")
+                axis >= relay.const(0, "int64"), axis, axis + relay.const(rank_output, "int64")
             )
             result = _op.expand_dims(result, axis)
         return result

tests/python/frontend/onnx/test_forward.py

@@ -676,6 +676,45 @@ def test_unsqueeze(target, dev):
     verify_with_ort(model, [in_shape], target=target, dev=dev, opset=11)
 
 
+@tvm.testing.parametrize_targets
+def test_unsqueeze_with_neg_axes(target, dev):
+    def verify_unsqueeze_with_neg_axes(opset=11):
+        in_shape = (2, 3, 4)
+        axis = (-2, -1)
+        out_shape = (2, 3, 4, 1, 1)
+        if opset < 13:
+            y = helper.make_node("Unsqueeze", ["in"], ["out"], axes=list(axis))
+            nodes = [y]
+        else:
+            axes = np.array(list(axis)).astype(np.int64)
+            axes = helper.make_node(
+                "Constant",
+                inputs=[],
+                outputs=["axes"],
+                value=onnx.helper.make_tensor(
+                    name="const_axes",
+                    data_type=onnx.TensorProto.INT64,
+                    dims=axes.shape,
+                    vals=axes.flatten().astype(int),
+                ),
+            )
+            y = helper.make_node("Unsqueeze", ["in", "axes"], ["out"])
+            nodes = [axes, y]
+
+        graph = helper.make_graph(
+            nodes,
+            "squeeze_test",
+            inputs=[helper.make_tensor_value_info("in", TensorProto.FLOAT, list(in_shape))],
+            outputs=[helper.make_tensor_value_info("out", TensorProto.FLOAT, list(out_shape))],
+        )
+
+        model = helper.make_model(graph, producer_name="squeeze_test")
+        verify_with_ort(model, [in_shape], target=target, dev=dev, opset=opset)
+
+    verify_unsqueeze_with_neg_axes()
+    verify_unsqueeze_with_neg_axes(opset=13)
+
+
 @tvm.testing.parametrize_targets
 def test_gather(target, dev):
     """test_gather"""